Abstract

Neural simulation-based inference is a recent paradigm shift in statistical science that uses simulations from a statistical model to train a neural network for statistical inference. These neural models serve as fast and accurate surrogates for classical inferential quantities, such as point estimators, confidence sets, likelihoods, posteriors or conditional distributions, whose exact evaluation tends to be intractable for many complex statistical models. In this talk, I will describe our recent work on developing a neural inference pipeline for complex spatial processes. In the first step of the pipeline, we infer the process parameters using a neural surrogate model for the likelihood function, and in the second step, we perform spatial prediction by training a neural diffusion model for spatial conditional simulation. The entire pipeline is amortized and only requires unconditional simulations from the spatial model during training. As such, this approach holds the potential for massive scale-up and increased fidelity of spatial inference in complex real-world settings that have previously relied on crude approximations or overly simplified modeling.